Validation of an AI‑Assisted Open‑Source software (CoreSlicer) for CT‑Based body composition analysis in locally...

This study aimed to validate the agreement between CoreSlicer (AI-assisted segmentation with manual calibration) and SliceOmatic (manual segmentation) for CT-based body composition analysis in gastric cancer patients.

This retrospective study included 261 patients with locally advanced gastric adenocarcinoma. Pre- and post-treatment L3-level CT images were analyzed using both software programs. Skeletal muscle area (SMA), visceral adipose tissue area (VAT), and subcutaneous adipose tissue area (SAT) were measured, and corresponding indices (SMI, SAI, VAI) were calculated. Agreement was assessed using intraclass correlation coefficients (ICC), Bland-Altman analysis, and Cohen's kappa. Multivariable logistic regression models were constructed to predict major pathological response (MPR), and model discrimination was compared using the area under the curve (AUC).

Excellent agreement was observed for SMA (ICC: pre-treatment 0.997, post-treatment 0.997), SAT (ICC: 0.990, 0.990), and VAT (ICC: 0.964, 0.942). Bland-Altman analysis showed minimal bias for SMA and SAT, whereas VAT was systematically overestimated by CoreSlicer. Diagnostic agreement for muscle atrophy was excellent (kappa: 0.923). Both methods showed similar predictive performance for MPR (AUC = 0.675).

CoreSlicer with manual calibration demonstrates excellent agreement with SliceOmatic for body composition measurements, supporting its use as a reliable alternative to manual segmentation in clinical practice.